🔥較早收集於 13m

SK 海力士與閃迪啟動 HBF 下一代儲存器全球標準化

SK 海力士與閃迪啟動 HBF 下一代儲存器全球標準化
PostLinkedIn
🔥閱讀原文: 36氪
#memory-standard#ai-inferencehbfsk-hynixsandiskhbf

💡New HBF memory standard for AI inference; key for efficient scaling

⚡ 30-Second TL;DR

有什麼變化

啟動 HBF 標準化聯盟

為什麼重要

推動高頻寬儲存器標準進展,對擴展 AI 推理至關重要,有助資料中心效率提升。

下一步行動

Review HBF specs on SK Hynix site for AI hardware integration planning.

誰應關注:Developers & AI Engineers

關鍵要點

  • 啟動 HBF 標準化聯盟
  • 針對 AI 推理工作負載
  • 於閃迪米爾皮塔斯總部聯合舉辦
  • SK 海力士與閃迪合作

🧠 深度解析

背景與延伸:來自公開資料,非原文內容。引用 9 個來源。

🔑 增強重點摘要

  • HBF positions as a new memory layer between HBM and SSD, offering HBM-comparable bandwidth with 8-16x greater capacity at similar cost for AI inference.[1][2][4]
  • SanDisk's HBF leverages advanced BiCS NAND technology and proprietary CBA wafer bonding, with a first-generation prototype featuring 16-layer chips up to 512GB per stack.[2][4][5]
  • The collaboration includes a dedicated workstream under the Open Compute Project (OCP) and follows an MOU signed prior to the kick-off event.[1]
  • SanDisk formed a Technical Advisory Board with industry experts, including input from AI players and figures like Raja Koduri, and won 'Best of Show, Most Innovative Technology' at FMS 2025.[2][3][4]

🛠️ 技術深入

  • HBF uses SanDisk’s BiCS 3D-NAND technology with proprietary CBA (Chip-on-Wafer Bonding) wafer bonding to integrate NAND flash into HBM-like packages.[2][4]
  • First-generation HBF features 16-layer memory chips, enabling up to 512GB capacity per stack while matching HBM bandwidth.[5]
  • Architecture breaks NAND memory arrays into smaller sections operating simultaneously, stacked vertically for higher capacity (8-16x HBM) at comparable cost and power efficiency.[5]
  • SK hynix's related 'LPW NAND' trademark enhances I/O channels while reducing per-channel speed to lower power for AI inference.[5]

🔮 前景展望AI analysis grounded in cited sources

SanDisk targets HBF samples in 2H 2026
Press releases confirm first HBF memory samples planned for second half of 2026, with AI-inference devices in early 2027.[2][3][7]
HBF standardization under OCP to drive ecosystem adoption
Joint workstream launched under Open Compute Project will define specs and foster industry-wide compatibility for AI infrastructure.[1]
HBF reduces AI data center TCO via capacity and efficiency
Positioned between HBM and SSD, HBF bridges performance-capacity gaps to lower total cost of ownership for inference workloads.[6]

時間線

2025-07
SanDisk forms Technical Advisory Board for HBF development.
2025-08
SanDisk unveils HBF prototype at FMS 2025, wins Most Innovative Technology award, signs MOU with SK hynix.
2025-08
SK hynix files 'LPW NAND' trademark related to high-performance inference memory.
2026-02
SK hynix and SanDisk host HBF Spec Standardization Consortium kick-off at SanDisk Milpitas HQ.
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源: 36氪

這是摘要,不是原文。去看原站,或訂閱每週簡報。

每週 AI 簡報

每週一封,可隨時退訂。